TECHNICAL MANUAL

Æon Documentation

The definitive guide to building dependable LLM agents with chaos engineering at the core.

Introduction

Æon is not an LLM wrapper. It is a chaos-native runtime that wraps agent reasoning in a deterministic Safety Envelope. Most frameworks are Think → Act. Æon is Think → Verify → Act.

WHY ÆON

Verification happens before execution. That makes Æon suitable for high-consequence environments where an unchecked action is unacceptable.

Installation

Æon requires Python 3.10+ on Linux, macOS, or WSL.

BASH

# 1 · install the resilience runtime
pip install aeon-core

# 2 · verify the toolchain
aeon doctor

Quickstart

An agent declares its Chaos Budget up front, the envelope it promises to stay inside.

PYTHONagent.py

from aeon import Agent, ChaosBudget

agent = Agent(
    name="sentinel",
    model_provider="ollama",
    model_name="mistral",
    # the Safety Envelope, declared up front
    chaos_budget=ChaosBudget(p99_ms=300, error_rate=0.02),
)
agent.start()

Architecture

The runtime coordinates seven interdependent layers as a closed cybernetic cycle. Observability is separated from execution (COM/MON).

Axioms

Axioms are deterministic Python functions that intercept actions before execution. Guardrails with teeth, not prompts.

PYTHONDEFINING AN AXIOM

from aeon.executive import axiom

@axiom(on_violation="BLOCK", response_time_ms=0.1)
def envelope_guard(metric):
    # deterministic kill-switch, runs before any action
    if metric.p99_deviation > metric.budget:
        return False   # BLOCK → Safe State
    return True        # GO

Chaos experiments

Inject faults into agent runs and tool calls inside a bounded blast radius, then watch the axioms respond.

BASH

# inject a fault inside a bounded blast radius
aeon chaos run "degrade gateway latency" --blast-radius=staging

# → [L4] fault injected · ALERT → LIMIT → BLOCK · Safe State restored

Resilience gate (CI/CD)

Chaos becomes a mandatory pipeline step. A version is promoted only if it holds the Safety Envelope under induced failure.

YAML.gitlab-ci · RESILIENCE GATE

stages: [deploy-staging, chaos-test, slo-validation, deploy-production]

chaos-test:
  script:
    # latency injection in the gateway service
    - kubectl apply -f chaos/latency-injection.yaml

slo-validation:
  script:
    - P99=$(promql "histogram_quantile(0.99, http_latency)")
    - ERR=$(promql "rate(http_errors[1m])")
    # gate on the Safety Envelope → implicit BLOCK
    - if [ "$P99" -gt "300ms" ] || [ "$ERR" -gt "0.02" ]; then
        echo "SLO violation, BLOCK"; exit 1; fi

Audit & provenance

Every reasoning step, tool call and axiom decision is chained into an immutable trail, recorded asynchronously, off the critical path.

LOGprovenance.chain

# every decision is chained, off the critical path
event=AXIOM_BLOCK  ts=2026-06-18T15:58:13Z
  prev=0x7c1a…b4  hash=0x9f3a…e1  sig=ed25519:ok

CLI reference

The aeon command manages the full lifecycle.

aeon init <name>             # scaffold a resilient agent
aeon serve --port 8000      # start the gateway
aeon chaos run "<goal>"      # run a controlled fault
aeon audit verify           # replay the provenance chain

Configuration

A single aeon.yaml declares the agent, its budget, and audit guarantees.

YAMLaeon.yaml

agent:
  name: "sentinel-01"
  trust_level: "restricted"  # full · restricted · isolated
chaos_budget:
  p99_ms: 300
  error_rate: 0.02
audit:
  mode: "off-chain-buffer"
  finality: "deterministic"